1. Path Diversity for Media Streaming The Use of Multiple Description Coding J. Apostolopoulos, M. Trott and W. Tan Presented by Xiaoyuan GUO

2. Outline  Introduction  Path Diversity  Multiple Description Coding  Systems Design, Analysis and Operation  Applications and Architectures  Summary 2/44 Outline 2016-2-27

3. Background and Motivation  Media Streaming Characteristics  Delay sensitive  Loss tolerant  Best-effort packet networks(i.g. Internet)  Delay, loss rate and available bandwidth dynamic and unpredictable  The default path is often not the best path  many applications(e.g. multicast or broadcast) lack a back channel or other means for retransmission  Media streaming is quite challenging 2016-2-27 Introduction 3/44

4. Solution  Path diversity + Multiple Description Coding 2016-2-27 Introduction 4/44

5. Outline  Introduction  Path Diversity  Multiple Description Coding  Systems Design, Analysis and Operation  Applications and Architectures  Summary 2016-2-27 Introduction 5/44

6. Definition and Categories  Diversity: frequency, time and spatial  A transmission technique that sends data through two or more paths in a packet-based network  Categories  Transmission methods  Use multiple paths at the same time  Select best path  Source numbers 2016-2-27 Path Diversity 6/44

7. Single Source  Path diversity using single source 2016-2-27 Path Diversity 7/44

8. Multiple sources  Path diversity using multiple sources 2016-2-27 Path Diversity 8/44

9. Benefits  Benefits of path diversity  Bandwidth aggregation(use all at once)  Traffic load balancing  Reduce probability of service outage, i.g. streaming from multiple servers in CDN or from multiple peers in P2P 2016-2-27 Path Diversity 9/44

10. Reduce delay and jitter  Reduce delay and delay jitter  Queue diversity  Network delay due to backlogged queue  Benefits of multiple parallel queue  Helpful  Time-invariant paths having different but unknown delay characteristics  Applications of end-to-end delay constraint while maintaining quality 2016-2-27 Path Diversity 10/44

11. Reduce loss  Reduce losses  Path selection  When a single path system is unreliable and with feedback 2016-2-27 Path Diversity 11/44

12. Reduce Uncertainty  Reduce uncertainty with averaging loss rate  Time-invariant paths lack feedback  Time-varying paths that measurement lags variation  Broadcast or multicast scenarios 2016-2-27 Path Diversity 12/44

13. Reduce Loss  Decrease the probability of outage  Outage: all communication along a network path is lost for a sizable length of time  Single path p while two paths p^2  Reduce the length of burst losses  Increase the interpacket space on each path 2016-2-27 Path Diversity 13/44

14. Outline  Introduction  Path Diversity Multiple Description Coding  Multiple Description Coding  Systems Design, Analysis and Operation  Applications and Architectures  Summary 14/44 Outline 2016-2-27

15. What is it?  Descriptions: sets of compressed data  The more descriptions available, the better the quality of the reproduction 2016-2-27 Multiple Description Coding 15/44

16. Why Use MD?  SD vs. SC vs. MD 2016-2-27 Multiple Description Coding 16/44 SD (single description) SC (Scalable coding) MD (multiple des.) No scalable No priority High code efficiency Scalable Priority to base layer Acceptable code efficiency Scalable No priority Acceptable code efficiency

17. Why Use MD?(2)  Best-effect network  Packets equally to be lost or delayed  Path diversity system  Send descriptions over different paths  Enhance the benefits of multiple paths 2016-2-27 Multiple Description Coding 17/44

18. How Good MD Can Do?  Rate-distortion theory  gives theoretical bounds for how much compression can be achieved using lossy compression methods  Rate-distortion function D(R) give the min. achievable average distortion D when source is described using R bits per source symbol  Gaussian source D(R)=2^(-2R) 2016-2-27 Multiple Description Coding 18/44

19. R/D for Multiple Description  Rate-distortion for multiple description  Single- and multiple-description upper bound can not be achieved simultaneously  Trade-off curve is known only for Gaussian source  Refer to L. Ozarow. “On a Source Coding Problem with Two Channels and Three Receivers,”  No obstacle to practical exploration  Growing literatures on practical schemes 2016-2-27 Multiple Description Coding 19/44

20. MD Speech and Audio Coding  Sequence of packets and sensitive to loss  Frame  even/odd samples  coded and sent in separate packets  Path diversity: Independently of the neighboring frame  reduce burst loss 2016-2-27 Multiple Description Coding 20/44

21. MD Image Coding  Source coding + FEC  MD-FEC  Scalable, prioritized  nonprioritized 2016-2-27 Multiple Description Coding 21/44 subsampling transform coding

22. MD-EFC 2016-2-27 Multiple Description Coding 22/44

23. MD-FEC(2)  Flexibility  Any number of scalable layers  Various amounts of FEC per layers  Arbitrary number of descriptions  Refer to R. Puri and K. Ramchandran. “Multiple Description Source Coding Using Forward Error Correction Codes” 2016-2-27 Multiple Description Coding 23/44

24. MD Video Coding  Consecutive video frames similar  Predictive coding  Two design extremes  Maintain good quality even when half data is lost  Accomplished independent prediction loops  Maximize coding efficiency  Single prediction loop  Mismatch with subsequent error propagation  Trade-off between coding efficiency and resilience to full or partial loss of stream 2016-2-27 Multiple Description Coding 24/44

25. MD Video Coding(2)  Refer to B. Heng, J. G. Apostolopoulos, and J. S. Lim. “End-to-End Rate-Distortion Optimized MD Mode Selection for Multiple Description Video Coding,”  Small number of description streams  Scalable MD coding(i.g. MD-FEC)  Large number of description streams 2016-2-27 Multiple Description Coding 25/44

26. Repairable MD Coding  Error propagation in predictive coding  Repairable MD coding attempt to stop error propagation  repairs one description using uncorrupted frames from the other description as long as not all the descriptions are simultaneously lost  Complemented with path diversity 2016-2-27 Multiple Description Coding 26/44

27. Repairable MD Coding 2016-2-27 Multiple Description Coding 27/44

28. Outline  Introduction  Path Diversity  Multiple Description Coding  Issues on Systems Design, Analysis and Operation  Applications and Architectures  Summary 28/44 Outline 2016-2-27

29. Issues on System Design  Joint and disjoint paths  Multiple paths maybe independent  bottlenecks occur on shard portion impact  Identify bottlenecks and avoid them  How many paths to use  Specifics of the application  Repairable MD coding, two paths enough  Benefits that one is trying to exploit  number of paths increase, aggregated bandwidth increases, probability of outage decrease and delay variability decrease 2016-2-27 System Design, Analysis and Operation 29/44

30. Issues on System Analysis  Characteristics of the available paths  Select the best paths or best servers  Min. distance while max. path diversity  Model path diversity performance  Goal  Select the best subset of possible servers  Select the best subset of paths  Compare path diversity scenarios  One distortion model for MD with two path diversity 2016-2-27 System Design, Analysis and Operation 30/44

31. State Transition Pro. Model 00: two descriptions are correctly received 01/10: one description is correctly received 11:two descriptions are lost simultaneously 2016-2-27 System Design, Analysis and Operation 31/44

32. One Path Diversity Model  Simplify network topologies  Refer to J. G. Apostolopoulos, W. Tan, S. J. Wee, and G. W. Wornell. “Modeling Path Diversity for Multiple Description Video Communication,” 2016-2-27 System Design, Analysis and Operation 32/44

33. Issues on System Operation  Streaming and packet scheduling across asymmetric paths  Path measurement  Estimate the characteristics of each path, i.g. time- varying, available bandwidth, etc.  Rate adaption  Packet scheduling  More important packet may be sent over the path with lower packet loss rate 2016-2-27 System Design, Analysis and Operation 33/44

34. Outline  Introduction  Path Diversity  Multiple Description Coding  Systems Design, Analysis and Operation  Applications and Architectures  Summary 34/44 Outline 2016-2-27

35. Low-delay application  Low-delay applications(i.g. VoIP)  Choose a low-latency path  Skype using nodes in internet to bypass firewall and NATS 2016-2-27 Applications and Architectures 35/44

36. CDN  Using multiple paths simultaneously  Content delivery networks  Choose a good/best server by multiple path selection  Multiple servers provide path diversity with SD and FEC  Multiple resources stream packets to client  MD-CDN design and operation  Refer to W. Tan et al. “On Multiple Description Streaming with Content Delivery Networks” 2016-2-27 Applications and Architectures 36/44

37. P2P  Peer-to-Peer networks  One essential challenge  Provide an uninterrupted flow of data to each client for the duration of the streaming session  Multiple servers reduce service outages in the face of uncertain peer and network condition 2016-2-27 Applications and Architectures 37/44

38. Wireless network  Path diversity over wireless network  Characteristics of wireless network  Time-varying and unpredictable behavior caused by multiple users, interface, propagation effects and mobility  Reduce uncertainty and improve reliabilities 2016-2-27 Applications and Architectures 38/44

39. Architectures  Control packet routes  Relay host  Direct different streams over different paths to send each stream to different relay hosts  Forward streams to their final destinations  Source routing  Specify the set of nodes for each packet to traverse  Specify different source routes for different subsets of packets 2016-2-27 Applications and Architectures 39/44

40. Outline  Introduction  Path Diversity  Multiple Description Coding  Systems Design, Analysis and Operation  Applications and Architectures  Summary 40/44 Outline 2016-2-27

41. Summary  Provide a survey of the benefits, architecture, system design issues and open problems associated with streaming delivery using path diversity  Path diversity  Overcome dynamic and unpredictable available bandwidth, delay and loss rate  Take a step closer to feedback-free video streaming  VoIP, CDN, P2P,WLAN, ad-hoc, etc. 2016-2-27 Summary 41/44

42. Summary(2)  Multiple description coding  Combined with path diversity to enhance its benefits for media streaming 2016-2-27 Summary 42/44

43. References J. G. Apostolopoulos, W. Tan, S. J. Wee, and G. W. Wornell. “Modeling Path Diversity for Multiple Description Video Communication,” IEEE ICASSP, May 2002. R. Puri and K. Ramchandran. “Multiple Description Source Coding Using Forward Error Correction Codes,” IEEE Asilomar Conference on Signals, Systems, and Computers, October 1999. J. G. Apostolopoulos, W. Tan, S. J. Wee, and G. W. Wornell. “Modeling Path Diversity for Multiple Description Video Communication,” IEEE ICASSP, May 2002. L. Ozarow. “On a Source Coding Problem with Two Channels and Three Receivers,”Bell Syst. Tech. J., 59:1909–1921, December 1980. J. G. Apostolopoulos, T. Wong, W. Tan, and S. J. Wee. “On Multiple Description Streaming with Content Delivery Networks,” IEEE INFOCOM, June 2002. 2016-2-27 Summary 43/44

44. 44/44 Outline 2016-2-27 Questions and Comments? Thank you